GAN Based Channel Estimation and Interference Cancellation for Mixed RF/FSO Cooperative Relaying Systems

Abstract

Free space optical (FSO) communication systems have been considered as a valid option to be integrated with the existing hardware of radio frequency (RF) systems in cooperative relaying technique called the dual hop RF/FSO asymmetric relay. However, the advantages of cooperative relaying is strongly hampered due to the presence of interference in practical systems. Moreover, the modelling of FSO systems need proper attention due to complex nature of the atmospheric turbulence, pointing errors and beam wander which involves tedious mathematical computations. The present research work focuses on the capability of deep learning models for developing a framework for enhancing the performance of the mixed RF/FSO relaying systems using interference cancellation. Specifically, generative adversarial networks (GANs) have been trained to represent the channel conditions of the co-channel interferer (CCI) and FSO links. The GAN model consists of two competing deep learning models — the generator and the discriminator. Two separate deep learning GAN models namely: CCI-GAN and FSO-GAN have been trained for estimating the channel state information (CSI) for the interferer and FSO links, respectively. The CCI-GAN model has been trained offline and tested with mean square error (MSE) of −53.01 dB and root mean square error (RMSE) of −11.003 dB. On the other hand, the FSO-GAN model has been trained with error metrics of MSE of -80.14 dB and RMSE of −11.12 dB. The trained models have been utilized to summarize the performance of the decode-and-forward (DF) mixed RF/FSO relaying network by analysing the outage probability, bit error rate (BER) and the ergodic capacity of the system. The obtained results demonstrate that the FSO-GAN generates the samples from the probability density function (PDF) representing the FSO link. More importantly, the interference cancellation mechanism adopted using the CCI-GAN enhances the system performance.